Spoken Language Recognition in the Latent Topic Simplex
Résumé
This paper investigates the use of latent topic modeling for spoken language recognition, where a topic is defined as discrete distribution over phone n-grams. These latent topics are trained in an unsupervised manner following the latent Dirichlet allocation (LDA) approach. We analyzed and observed that language cues can be represented very accurately in terms of the latent topics, where some latent topics are language specific while others exhibit multilingual characteristic. We then show that the latent topics can be used to define a low dimensional simplex (i.e., a bounded linear manifold) where language recognition can be done effectively. Experiments carried on the NIST 2007 language detection tasks show promising results.
Origine : Fichiers produits par l'(les) auteur(s)
Loading...